@ls-stack/agent-eval 0.40.0 → 0.42.0

package/dist/{runner-njK_CtXC.mjs → runner-B1Cyevvr.mjs}

@@ -1,2 +1,2 @@
-import { n as initRunner, t as getRunnerInstance } from "./runner-Bb9JdFkg.mjs";
+import { n as initRunner, t as getRunnerInstance } from "./runner-BG0L4yId.mjs";
 export { getRunnerInstance, initRunner };

package/dist/{runner-Bb9JdFkg.mjs → runner-BG0L4yId.mjs}

@@ -1,5 +1,5 @@
-import { n as createRunner } from "./cli-Sg8UdOnm.mjs";
-import "./src-BZzPFS8r.mjs";
+import { n as createRunner } from "./cli-Bk5g-bat.mjs";
+import "./src-t6OVp1li.mjs";
 //#region ../../apps/server/src/runner.ts
 let runnerInstance = null;
 function getRunnerInstance() {

package/dist/src-t6OVp1li.mjs

@@ -0,0 +1,13 @@
+import { Lt as defineEval$1, rt as matchesEvalTags$1 } from "./runOrchestration-DhTiT4V0.mjs";
+import "./cli-Bk5g-bat.mjs";
+//#region src/index.ts
+/** Register an eval definition with typed tag support. */
+function defineEval(definition) {
+	defineEval$1(definition);
+}
+/** Return whether the active eval case has tags matching the typed input. */
+function matchesEvalTags(input) {
+	return matchesEvalTags$1(input);
+}
+//#endregion
+export { matchesEvalTags as n, defineEval as t };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ls-stack/agent-eval",
-  "version": "0.40.0",
+  "version": "0.42.0",
   "type": "module",
   "bin": {
     "agent-evals": "./dist/bin.mjs"
@@ -33,8 +33,8 @@
     "@types/node": "^24.7.2",
     "typescript": "^5.9.2",
     "@agent-evals/runner": "0.0.1",
-    "@agent-evals/sdk": "0.0.1",
-    "@agent-evals/shared": "0.0.1"
+    "@agent-evals/shared": "0.0.1",
+    "@agent-evals/sdk": "0.0.1"
   },
   "scripts": {
     "build": "pnpm --filter @agent-evals/web build && pnpm --filter @agent-evals/shared build && pnpm --filter @agent-evals/sdk build && pnpm --filter @agent-evals/runner build && tsdown --filter cli-js && tsdown --filter cli-types",

package/skills/agent-eval/SKILL.md

@@ -25,7 +25,8 @@ display rules), read the TypeScript declarations shipped with the package:
 - The CLI automatically loads `.env` from the current workspace. Shell-provided
   environment variables win; pass `--no-env` to disable `.env` loading once.
 - Unfiltered `agent-evals run` is disabled by default; use `--eval` or `--case`
-  for targeted CLI runs. Set `allowCliRunAll: true` in
+  for targeted CLI runs, or `--tags-filter <expr>` to run cases matching tags.
+  Set `allowCliRunAll: true` in
   `agent-evals.config.ts` to opt into run-all CLI behavior. The web UI can
   still run grouped evals and confirms before starting more than five. On a
   single eval page, the Run chevron can open a picker to run specific authored
@@ -69,6 +70,25 @@ during case-owned phases by default; log arguments are stored as JSON-safe
 values and rendered with the JSON viewer, collapsed previews include best-effort
 code locations when stack data is available, previews are capped, and logs
 inside cached operations are not replayed from cache hits.
+Use eval tags to target related coverage without naming every case:
+`AgentEvalsConfig.tags` applies workspace-wide tags, `defineEval({ tags })`
+adds eval tags, `case.tags` adds case-only tags, and `removeTags` disables a
+configured global tag for one eval. CLI filters support Vitest-style tag
+expressions such as `agent-evals run --tags-filter "refunds && !slow"`.
+Inside eval-scoped code, use `matchesEvalTags('tag')` or
+`matchesEvalTags({ all, any, not })`; it uses typed exact tag names and returns
+`false` outside a case scope. Projects can narrow tag names with a `.d.ts`
+module augmentation:
+
+```ts
+import '@ls-stack/agent-eval';
+
+declare module '@ls-stack/agent-eval' {
+  interface AgentEvalTagRegistry {
+    tags: 'refunds' | 'media' | 'manual' | 'slow';
+  }
+}
+```
 
 ### Product code (instrumented once, reused everywhere)
 
@@ -333,6 +353,9 @@ See `EvalScoreDef` / `EvalManualScoreDef` in the types for the full shape
   column from the runs table when every rendered row is missing the value,
   `null`, or an empty string; `0` and `false` still count as values, and the
   value remains available in case details and raw output data.
+  In the case detail Output tab, string outputs that look like Markdown render
+  as Markdown even without `format: 'markdown'`, with a Preview/Raw toggle for
+  inspecting the original text.
 - `deriveFromTracing` can be authored globally in `agent-evals.config.ts` or
   locally on one eval. Prefer the keyed map form for shared metrics:
   `deriveFromTracing: { toolCalls: ({ trace }) => trace.findSpansByKind('tool').length }`.